Mia TensorFold runtime: functional validation — 2026-11-13¶
The pinned Mia update fixes a reproduced truncated-tool failure and expands usable prompt scoring. It is accepted as a same-model runtime candidate after direct regression checks and a verified retained-runtime rollback rehearsal. Operational promotion and fleet closure are tracked separately; this finding establishes no speed improvement.
Evidence index · Decision summary · Hashed artifact manifest.
What changed¶
Mia recipe fc1ee3a applies 110 author patches to TensorFold f350cd3e. The six owned telemetry patches produce tree 109491764fc5f711705c148da278a708f6c1aeea; the ordered portable catalog is in PR 755. Four owned patches are unchanged, and two were rebased around checkpoint identity fields. Authorship and notices remain intact.
The additions cover tool-envelope guards, truncated JSON tool calls and Messages conversion, additional EOS handling, verification windows up to32, actual-KV-pool reply caps, prompt scoring up to128 alternatives with requested token IDs, and build identity. Prefill returns to2048 rows. L2 prefetch and thinking-off splitting remain disabled.
Matched functional observations¶
The same weights and DFlash2 draft ran on two RTX PRO 6000 Blackwell Max-Q96GiB cards at300W, with TP2, EXL3 4-bpw, FP8 sharded KV, configured327680 context,32768 output,C4 and image8. Container memory was bounded at48GiB with zero effective swap and a16GiB host reserve. Separate conversation-cache volumes prevented cross-build cache reuse.
| Request or gate | Previous runtime | Updated runtime | Comparison |
|---|---|---|---|
| Forced write_file call cut at64 output tokens, nonstreamed Messages API | HTTP500, unterminated argument string | HTTP200, parsed partial input and max_tokens stop reason | matched request; concrete reliability fix |
| Prompt logprobs128 | HTTP400, cap20 | 128 accepted;129 refused | matched request; added capability |
| Requested IDs outside the top128 | unsupported in prior contract | both returned with distinct finite scores/ranks | partially matched; candidate-only scoring checks |
| Repeated and concurrent scoring | not collected | exact repeat; two scoring plus one plain request pass | candidate-only regression |
| OpenAI whole/stream and Messages stream truncated calls | valid cut-call response | valid cut-call response | matched structural assertions |
| Tools, tool results, streaming, JSON, image | retained prior qualification | native regression passes, including20 tool cases | regression, not a new broad quality score |
The synthetic tool was never executed. Tests verify transport and argument handling, not completion of a1000-number file. The requested-ID control catches reuse of the first requested ID's score. Numerical agreement with a full-precision reference was not measured.
Context, concurrency and recovery¶
Four concurrent exact canaries passed. The telemetry window retained1164 coherent samples, observed four active streams and no cgroup memory events; minimum observed host-available memory was68.04GiB. This does not prove four simultaneous full327680-token windows.
Exact retrieval passed at19456 and99284 prompt tokens. The first99K request already reused18432 tokens, so it is not a fully cold measurement. Warm requests reused18432/98304 tokens; after managed recreation, the same-volume first retrievals reused those prefixes and returned correct values. A384-entry structured map also passed.
The previous container was restored with its exact image and identity, monitoring was rebound to its new start time, and authenticated routed RESTORE_OK passed. The later delta window reused the accepted candidate. Restoration evidence distinguishes this rehearsal from the isolated candidate state recorded when the qualification packet was finalized.
Freshness and boundaries¶
Mia main was rechecked on November13 and remained fc1ee3a. The open video assembly PR reports results on a different Max-Q/CPU setup; it needs a separate WITH_VIDEO=1 image and remains outside this text/image route. Video is not advertised. External performance figures are priors, not local measurements.
Source validation passed1049 upstream host tests,60 telemetry tests and4548 product tests (116 skipped,20 subtests); all13 PR checks passed. The artifact set retains the baseline failures and exact request responses. Public copies replace private paths, topology and ephemeral identities with synthetic values.
Tool-guard history smoke does not guarantee rejection-path activation; host parser tests cover that path. No speed ranking, endurance, reboot, fresh-host rebuild, or universal quality claim follows. Full rollout closure remains harness_convergence_pending while the fleet's replacement workflow owner is unavailable; legacy writers stay retired.